How much water does AI use?
According to the most-cited research, running 20 to 50 questions through a cloud AI evaporates roughly half a litre of fresh water in data-centre cooling. Training GPT‑3 alone consumed about 700,000 litres.
But be careful with these numbers: estimates vary enormously depending on what's counted, where the data centre is, and what season it is. Anyone quoting a single precise figure is oversimplifying.
Why does AI use water at all?
Servers get hot. Many data centres use evaporative cooling: clean fresh water is evaporated to carry heat away. That water isn't returned to the river — it leaves as steam.
There's a second, larger bucket most headlines ignore: the water consumed off-site, generating the electricity the data centre runs on. Whether a study counts that changes the answer by an order of magnitude.
The numbers
| What | Water | Source |
|---|---|---|
| ~20–50 ChatGPT queries | ≈ 500 ml | UC Riverside |
| Training GPT‑3 (~2 weeks) | ≈ 700,000 L | UC Riverside |
| Google's US data centres (2021) | ≈ 12.7 billion L | UC Riverside |
For scale: the 700,000 litres used to train GPT‑3 is roughly the water needed to manufacture 370 BMW cars, or about 320 Teslas.
An honest caveat. Not all researchers agree. Some analyses put a single ChatGPT conversation far lower — in the range of 10–25 millilitres — because they count only direct on-site evaporation, use newer efficiency figures, or model different data centres. The half-litre figure is the most widely cited; it is not the only one. The direction is clear and the order of magnitude is real, but treat any exact number with suspicion.
The bigger story is electricity
Water is downstream of power. Data centres already consume roughly 1.5% of the world's electricity, and the International Energy Agency expects that to roughly double by 2030, driven substantially by AI. More power means more heat, and more heat means more cooling.
What actually reduces it
Honest answer: mostly things you don't control — data centres siting in cooler climates, closed-loop cooling, cleaner grids, smaller and more efficient models.
The one thing you do control is where the model runs. A model running on your own laptop evaporates no data-centre cooling water. It draws a modest amount of electricity from a device you're already using. That's not a claim that local AI "saves the planet" — your electricity has its own footprint — but it does remove your queries from the cooling-tower equation entirely.
AI that runs on your computer
No data centre. No cooling tower. No server storing your conversations. GreenCube runs the model locally — chat, read documents, understand images, fully offline.
Get GreenCube — €9Sources
- UC Riverside — AI programs consume large volumes of scarce water (Shaolei Ren et al., "Making AI Less Thirsty")
- International Energy Agency — Energy demand from AI